| """ |
| Copyright (c) 2023, salesforce.com, inc. |
| All rights reserved. |
| SPDX-License-Identifier: Apache License 2.0 |
| For full license text, see the LICENSE file in the repo root or https://www.apache.org/licenses/LICENSE-2.0 |
| """ |
|
|
|
|
| import os |
| os.environ["OPENAI_API_KEY"] = "" |
|
|
|
|
| from langchain.chains.llm import LLMChain |
| from langchain.prompts import PromptTemplate |
| from langchain.chat_models import ChatOpenAI |
| import json |
| from utils import open_json, save_json, open_jsonl |
| from collections import defaultdict |
|
|
| class EvaluateDialogs(object): |
| """ Evaluate Dialogs based on OpenAI. To run this: |
| pip install openai |
| pip install langchain |
| """ |
| def __init__(self): |
| self.data_dir = "/Users/jianguozhang/TOD-Family/TOD-Studio/open-source/" |
| self.excluded_datasets = ['MetaLWOZ', "MuDoCo", "SalesBot", "HDSA-Dialog", "MULTIWOZ2_2"] |
|
|
| self.quality_agent_prompt = PromptTemplate( |
| input_variables=["dialog"], |
| template=""" |
| Hi AI, I plan to train a language model for response generation. Please analyze the following dialogue and evaluate it based on the criteria provided. Assign a score from 1 (poor) to 5 (excellent) for each category. We're looking for a critical assessment, and higher scores should only be given to truly exceptional examples. The criteria for evaluation are: Understanding, Relevance, Completeness, Correctness, and Coherence. |
| |
| After your assessment, provide an overall score for the dialogue along with a concise summary of your evaluation. The overall score should also be on a scale of 1 (poor) to 5 (excellent) and should represent a holistic assessment of the dialogue. |
| |
| Please present your evaluation and comment into the following format: |
| |
| {{ |
| "Understanding": _, |
| "Relevance": _, |
| "Completeness": _, |
| "Correctness": _, |
| "Coherence": _, |
| "Overall": {{"score": _, "comment": _}} |
| }} |
| |
| Please replace each underscore (_) with the appropriate score. For the 'Overall' field, provide the score and a concise comment. Regarding to the comment, it should not only summarize the dialogue's quality but also highlight any issues or shortcomings you may have identified in the dialogue. |
| |
| Below is the dialog: |
| |
| {dialog} |
| |
| Evaluate the dialog now. |
| """ |
| ) |
|
|
| self.quality_chain = LLMChain(llm=ChatOpenAI(temperature=0.2, model_name="gpt-3.5-turbo"), prompt=self.quality_agent_prompt) |
|
|
| def run_openai_evaluation(self, dialog): |
| res = self.quality_chain.run(dialog=dialog) |
| try: |
| res = json.loads(res) |
| except: |
| res = str(res) |
| return res |
|
|
| def tod(self): |
| """ |
| Evaluate TOD dialogues |
| :return: |
| """ |
| folder_name = "Task-Oriented-Dialogues--OpenAI" |
| folder_path = os.path.join(self.data_dir, folder_name) |
| dataset_names = os.listdir(folder_path) |
| print(dataset_names) |
| print() |
| for dataset_name in dataset_names: |
| if not os.path.isdir(os.path.join(folder_path, dataset_name)): |
| continue |
|
|
| data = open_json(os.path.join(folder_path, dataset_name, "train.json")) |
| f_writer = open(os.path.join(folder_path, dataset_name, "train_quality_scores.json"), "w") |
| print("Start processing: {} #total dialogs: {}".format(dataset_name, len(data))) |
|
|
| for index, item in enumerate(data): |
|
|
| output = defaultdict(dict) |
| output["source"] = item["source"] |
| output["quality score"] = self.run_openai_evaluation(item["dialog"]) |
|
|
| json.dump(output, f_writer) |
| f_writer.write("\n") |
| if index % 10 == 0 or index + 1 == len(data): |
| f_writer.flush() |
|
|
| def run(self): |
| self.tod() |
|
|
| process = EvaluateDialogs() |
| |
| process.run() |
|
|
